WO1994010682A1 - Method of encoding speech - Google Patents

Abstract

The invention concerns a speech-encoding method using analysis-by-synthesis techniques. The speech signal is scanned, a frame formed from a predetermined number of scanning samples, and the coefficients of a grade P speech-synthesis filter determined from the samples in each frame. Using these coefficients, a number P of so-called line-spectrum parameters (LSPs) are determined and quantized, for transmission over a channel with limited transmission capacity. The method is characterized in that every other line-spectrum parameter (LSP) is quantized in scalar (absolute) fashion and that the line spectrum parameters (LSPs) lying between them are transformed (normalized) before quantization. The invention is intended particularly for use in speech-encoding equipment in portable radio equipment.

Description

description

Method for speech coding The invention is based on a method for speech coding using the analysis-by-synthesis method according to the preamble of claims 1 and 2, respectively

 Speech coding methods are known, for example from German Patent 38 34 871.

The speech coding procedure has one thing in common

Prediction analysis of the input signal (linear prediction coder, LPC). The speech signal at the input of the

Encoders within a certain period of e.g. 20-30 ms divided. Each speech frame is subjected to a linear prediction analysis in the encoder, which removes linear dependencies in the speech signal. The linear prediction is carried out with the help of FIR filters (Finite Impulse Response). The coefficients of these linear filters are in each

Frame redetermined, i.e. these are adaptive filters.

 Today's speech coders which operate at bit rates between 4 and 16 kbit / sec. work, generally use the analysis-by-synthesis method, in which the filter coefficients listed above and an associated excitation are determined in the transmitter so that the energy of the weighted error e (n) between the original language and the synthesized language is as small as possible.

Parameters describing the excitation and the coefficients of the linear filter already mentioned above must be transmitted to the receiver. The determination of the coefficients of the linear filter will not be discussed in more detail here. As a result you get a Grade P non-recursive filter with the

Transfer function

The inverse transfer function H (z) = 1 / A (z) converts the spring signal (the excitation) into the (synthesized)

Speech signal at:

The filter H (z) calculated according to this method is without

Quantization of the filter coefficients a _i stable in any case

However, the filter coefficients a _i have a large dynamic range and are therefore poorly suited for quantization and transmission. Besides, there is no easy one

Possibility of the stability of the recursive in the receiver

Check filters.

 It is known that the so-called line spectrum parameters LSP for quantization and transmission

Description of the predictor filter H (z) are suitable. These parameters are obtained as zeros of a symmetric polynomial

F ₁ (z) = A (z) + Z ^{- (P + 1)} A (z ^-1 ) and an antisymmetric polynorm

F ₂ (z) = A (z) - Z ^{- (P + 1)} A (z ^-1 )

The zeros z _Oi of F ₁ and F ₂ have the following properties, all zeros are on the unit circle, so they are adequately described by specifying a phase _i - all zeros are simple

- On the unit circle there is an alternating zero of F ₁ and F ₂ .

dargestellt werden. FIG. 2 shows the zeros of F ₁ (z) and F ₂ (z) for the cases P = 6 and P = 5. All zeros z _i can by the arguments ω _i or by the derived frequency value

being represented.

Since the zeros occur in conjugate complex pairs and there are zeros at ± 1 in any case, the polynomials F ₁ and F ₂ are _i by specifying P values

 completely determined.

 According to the properties described above must apply

ω ₁ <ω ₂ <... <ω _P

 This monotony property is imperative for the recursive filter H (z) to be stable. This gives you a criterion for checking the stability of the filter.

When changing the characteristic of the spectrum of the

 Input signal, the distribution of individual LSPs changes significantly. 1 shows the distribution of the LSP for filter grade P = 10 as an example. In the upper picture, Fig. La, the input language is only low pass filtered, in the lower picture, Fig. Lb, IRS filtered (band limited) according to CCITT P.48.

A common method is the scalar quantization of each individual LSP, for example, in 4.8 kbit / sec. CELP speech codec according to the Federal Standard 1016 of the US Department of Defense the Line Spectrum Parameters scalar quantized with a total of 34 bits.

 It should be noted in the quantization that even after the

. Quantization the monotonicity property must be obtained for the recursive filter to be stable; ie the following must apply:

,

Since the value ranges of the quantizers for ω _i and ω _i +1 overlap, all are after the quantization of ω _χ

Quantization levels of ω _{i + 1} excluded, which violate this strict monotony (see Figure 3). Conversely, even after the quantization of ω _{i + l} , values from the

Quantizer no longer permissible for ω _i . This means that some of the bits that are available for the quantization of the parameters LSP are not fully used. According to FIG. 3 there are 8 possible steps for ω _{i + l}

 actually only 5 used.

Another disadvantage of this method is that adaptation to different input spectra of the speech signal is not possible. If the quantizer can be used for this, the range of values for individual line spectrum parameters increases. This leads to an increase in the bit rate.

References [5] and [6] suggest reducing the bit rate for the transmission of the line spectrum parameters by quantizing their differences. The first LSP is scalarized as above.

Für alle weiteren LSP wird die Differenz zum vorangegangenen Wert berechnet und diese dann quantisiert.

For all other LSPs, the difference to the previous value is calculated and then quantized.

 This method adapts well to different ones

Input spectra of the speech signal, since only the value range of the first LSP has to be chosen sufficiently large.

A disadvantage of this method is the propagation of errors. If an error occurs during the transmission of ω _x , all ω _i , for i = x to P, are decoded incorrectly.

The present invention was based on the object

Method of the type mentioned at the beginning, which is able to achieve an improvement in the speech quality with a constant bit rate or a reduction in the bit rate with a constant speech quality.

In addition, a reduction in the sensitivity of the

Speech codecs can be achieved compared to speech signals with different input characteristics. The one needed

Circuitry should not be too high.

This object was achieved by claims 1 and 2.

Advantageous configurations result from the

Dependent claims.

The method according to the invention achieves the advantages of an improvement in the speech quality with a constant bit rate or a reduction in the bit rate with a constant bit rate

Voice quality. In addition, the method according to the invention has a reduced sensitivity of the speech codec to speech signals with very different input spectra. Another advantage is that a Transmission error with an LSP only affects a maximum of two further LSP values.

The invention is based on the idea of neither quantizing all LSP parameters scalarly nor quantifying only a single one of the total P parameters scalarly, but rather only quantizing every nth of the P parameters scalarly and the in between

transform or map lying parameters and then quantize them.

 The process is described below using a

Embodiment described in more detail, wherein it is assumed that P is an even number.

In a first step, every second LSP becomes scalar

quantized.

Now must apply due to the strict monotony

where the fictitious value _{P + 1} to the maximum possible value for

und

. Ideal wäre es.nun, für jede Kombination von

und

einen eigenen _{P is} set. This range of values for ω _i changes from frame to frame

and

, It would be ideal for any combination of

and

an own

To use quantizers for ω _i , which is for the sake of

Jeder Wert X_i kann nun mit einem Quantisierer quantisiert und übertragen werden. Die Rücktransformation erfolgt gemäß

Realization effort is not possible. Instead, the value range is mapped to the interval] 0.1 [by the following transformation. )

Each value X _i can now be quantized and transmitted using a quantizer. The back transformation takes place according to

 The procedure works accordingly if you have the

 Parameters that are absolutely quantized with those

swapped, which are quantized according to normalization, ie quantize absolutely: ω _i i = 2, (2), P quantize after transformation: ω _i i = 1, (2), P - 1

Instead of transforming the LSP into the image area, it is also possible to map the quantizer from the image area according to (13) into the ω area.

Similarly, in the second embodiment, every third LSP is quantized scalarly.

für i = 1, (3),

for i = 1, (3),

The mapping function for the parameters in between are, for example

or

since ω _{i is} now known. This solution further reduces the

 Bit rate with the same quality or a higher one

 Quality at constant bit rate; however, a transmission error affects max. three more LSP values.

 A corresponding procedure can also be followed in which only every fourth LSP is scalarly quantified and those in between

lying LSP are transformed accordingly and then transmitted quantized.

literature

[1] Markel, J.D .; Gray, A.H .: Linear Prediction of Speech.

 Berlin, Heidelberg, New York: Springer Verlag, 1976

 [2] Müller, J.M .; Scheuermann, H .; Wächter, B .: A contribution to speech coding for bit rates below 8 kbit / s frequency, band

43, 9/89, pp.242-252

 [3] N. Sugamura, F. Itakura: "Speech Analysis and Synthesis

Methods Deveoped at ECL in NTT form LPC to LSP-. Speech

 Communication, Volume 5, 1986, pp. 199-215

 [4] J.P. Campbell, V.C. Welch, T.E. Tremain: "The DOD 4.8 kbps

Standard ", from" Advances in Speech Coding ", Kluwer, 1991

 [5] F.K. Soong, B.H. Juang: "LSP and Speech Data Compression";

Proc. ICASSP-84, March '84

 [6] F.K. Soong, B.H. Juang: "Optimal Quantization of LSP

 Parameters "Proc. ICASSP-88, April'88

Claims

Claims 1. A method for speech coding using the analysis-by-synthesis method, wherein the speech signal is sampled, a frame is formed from a fixed number of samples and the coefficients of the samples are frame-by-frame Speech synthesis filters with the degree P are determined, a number P of so-called line spectrum parameters LSP being determined and quantized by means of these coefficients, for transmission over a channel with limited Transmission capacity, characterized in that every second line spectrum parameter LSP scalar (absolute) is quantized

for i = 1, (2), P-1 or i = 2, (2), P and that the line spectrum parameters LSP ω _{i in between} for i = 2, (2), P or i = 1, ( 2), P-1 are transformed (normalized) before quantization.

 2. A method for speech coding using the analysis-by-synthesis method, wherein the speech signal is sampled, a frame is formed from a fixed number of samples and the coefficients of the samples are frame-by-frame Speech synthesis filters with the degree P are determined, a number P of so-called line spectrum parameters LSP being determined and quantized by means of these coefficients, for transmission over a channel with limited Transmission capacity, characterized in that everyone third line spectrum parameter LSP scalar (absolute) is quantized for i = 1, (3), ... or i = 2, (3), ... or i = 3, (3), ...

and that the intermediate line spectrum parameters LSP ω _i for i = 2, 3, 5, 6 ... or i = 1, 3, 4, 6, 7 ...  or i = 1, 2, 4, 5, 7, 8 ... with

transformed and then quantized. 3. A method for speech coding using the analysis-by-synthesis method, wherein the speech signal is sampled, a frame is formed from a fixed number of samples and the coefficients of the samples are frame-by-frame Speech synthesis filters with the degree P are determined, a number P of so-called line spectrum parameters LSP being determined and quantized by means of these coefficients, for transmission over a channel with limited Transmission capacity, characterized in that every nth line spectrum parameter LSP is scalar (absolute) quantized

for i = m, (n), P; l <m <n and that the lines between them transform spectrum parameters LSP ω _i for i = 1, P and i ≠ m, (n), P (normalized)

and then quantized. 4. The method according to claim 1, characterized in that the transformation according to the function

he follows. 5. The method according to claim 2, characterized in that the transformation according to the mapping functions

or

he follows.